OpenAI’s o1-pro, LG’s EXAONE Deep, and Hugging Face’s HuggingSnap iOS App

Summary

AI Deep Dive Podcast Summary

Podcast Title: AI Deep Dive
Host/Author: Daily Deep Dives
Episode Title: OpenAI’s O1-Pro, LG’s EXAONE Deep, and Hugging Face’s HuggingSnap iOS App
Release Date: March 20, 2025

Welcome to this detailed summary of the AI Deep Dive podcast episode, where hosts A and B explore the latest advancements and challenges in the artificial intelligence landscape. This episode delves into three significant developments: OpenAI's new O1 Pro model, LG's EXAONE Deep, and Hugging Face's innovative HuggingSnap iOS app. The discussion not only highlights the technological breakthroughs but also examines the ethical and practical implications of these AI advancements.

---

1. OpenAI’s O1 Pro Model

Introduction to O1 Pro

The episode begins with an exploration of OpenAI’s latest offering, the O1 Pro model. Hosts A and B discuss its positioning as a more powerful iteration of the existing O1 model, specifically tailored for developers requiring substantial computational resources.

A [02:04]: "Let's start with OpenAI and this new O1 Pro model. What is this thing?"

Features and Pricing

O1 Pro is described as a turbocharged version of the O1, boasting enhanced capabilities aimed at solving complex problems. However, this power comes at a premium price point:

B [02:18]: "We're talking 150 per million dollar input tokens and $600 per million dollar output tokens. Wow."

Compared to its predecessors, the O1 Pro's pricing is significantly higher, indicating OpenAI's target market of developers who prioritize performance over cost.

Performance and Reception

Despite its advanced specifications, early access feedback revealed some surprising shortcomings. Users reported that O1 Pro struggled with tasks that seemed straightforward, such as solving Sudoku puzzles and interpreting optical illusion jokes.

A [03:00]: "But here's where it gets interesting. ... people... they weren't blown away. They were saying it struggled with things like Sudoku puzzles and even some optical illusion jokes."

OpenAI claims that while the performance improvement over the regular O1 is modest, the O1 Pro offers greater reliability, which may justify the additional expense for certain applications.

B [03:31]: "They did say it was more reliable, so maybe that's the key."

Implications for Developers and the AI Market

The introduction of O1 Pro signals a trend toward increasingly specialized AI tools. Developers must now evaluate AI models not just on overall power but on specific use-case performance and reliability.

B [04:03]: "It's a good question. ... increasingly specialized AI tools. ... which AI is the right tool for this specific job."

Conclusion

OpenAI's O1 Pro model represents a significant investment for developers seeking robust AI capabilities. However, the mixed early feedback underscores the importance of balancing cost with tangible performance benefits.

---

2. Privacy Complaint Against ChatGPT

Overview of the Privacy Complaint

The conversation shifts to a serious privacy complaint filed against ChatGPT by the Norwegian privacy rights group, NOIB. The complaint centers on ChatGPT fabricating damaging false information about an individual.

A [04:36]: "It's about AI making things up. And it's not just harmless little errors. This time it accused someone of a very serious crime..."

Details of the Case

NOIB represents a Norwegian individual falsely accused by ChatGPT of murdering his children—a claim entirely untrue and potentially defamatory.

A [05:00]: "...Chatgpt accused him of, well, murdering his children, which, of course, he absolutely did not do."

NOIB’s Arguments and GDPR Concerns

NOIB argues that under the General Data Protection Regulation (GDPR), individuals have the right to correct inaccurate personal information. They assert that OpenAI has failed to provide adequate mechanisms for users to rectify these errors.

B [05:28]: "Nove is saying a simple disclaimer like, hey, ChatGPT can make mistakes. That's not enough. They are saying OpenAI is legally obligated to make sure the information they're putting out there is accurate."

The complaint highlights the severe implications of AI-generated misinformation, emphasizing the need for accountability and robust error-correction protocols in AI systems.

Regulatory Landscape and Potential Penalties

The discussion references previous regulatory actions, such as Italy's temporary ban and subsequent fines on OpenAI for GDPR violations, illustrating the stringent stance European regulators are taking.

B [05:50]: "Remember what happened in Italy last year? They temporarily blocked ChatGPT because of GDPR concerns..."

Potential penalties under GDPR can be substantial, reaching up to 4% of a company's global revenue, underscoring the financial risks for non-compliance.

Broader Implications for AI Accountability

The case exemplifies the critical need for AI developers to ensure the accuracy and reliability of their models. It raises important questions about the ethical responsibilities of AI providers in preventing and addressing harmful misinformation.

A [08:22]: "...highlight how important data privacy is in the age of AI. This is about protecting ourselves, but it's also about protecting others from being harmed by inaccurate information."

Conclusion

The privacy complaint against ChatGPT serves as a cautionary tale about the potential harms of AI inaccuracies. It underscores the imperative for AI companies to implement effective measures for data accuracy and user protection to comply with stringent data privacy regulations.

---

3. LG’s EXAONE Deep

Introduction to EXAONE Deep

The conversation progresses to LG’s latest AI innovation, EXAONE Deep, which marks a significant advancement in reasoning AI. LG positions this model as a new standard in the field, competing with major players.

A [08:44]: "It sounds like LG has really stepped up their AI game with this EXAONE Deep model."

Agentic AI and Capabilities

EXAONE Deep embodies the concept of agentic AI—systems capable of autonomous decision-making and problem-solving without constant human intervention.

B [08:56]: "It's the idea that AI systems can be more than just pattern recognizers. ... make decisions and solve problems without needing a human to hold its hand every step of the way."

Benchmark Performance and Efficiency

LG’s EXAONE Deep demonstrates exceptional performance across various benchmarks. Notably, models with 32 billion, 7.8 billion, and 2.4 billion parameters achieved top scores in the CSAT 2025 mathematics section and performed competitively in the American Invitational Mathematics Examination (AME).

A [09:21]: "...they all got the highest possible score on the CSAT 2025 mathematics section."

B [09:42]: "DeepSeek R1 is way bigger. It has 671 billion parameters. So Exxon Deep is doing more with less."

These results are particularly impressive given the model sizes, highlighting LG’s focus on efficiency and performance optimization.

Applications and Accessibility

LG aims to make AI more accessible by developing models that perform well on edge devices. The 2.4 billion parameter model is recognized as one of the best for edge applications, according to Hugging Face’s LLM Leaderboard.

B [10:34]: "It seems like LG is focusing on making AI that's not just powerful, but also accessible."

This approach ensures that advanced AI capabilities are available on more affordable and portable devices, broadening the scope of AI applications.

Endorsements and Recognition

EXAONE Deep has received significant recognition from Epoch AI, which highlighted both the 32B and EXOEN 3.5 models as notable, the only Korean models to receive such accolades in the past two years.

B [11:22]: "It's not just good at one thing, it seems to be good at, well, a lot of things."

This endorsement underscores LG’s achievement in developing versatile and high-performing AI models.

Conclusion

LG’s EXAONE Deep represents a leap forward in reasoning AI, combining high performance with efficiency. Its success on various benchmarks and recognition from industry leaders position LG as a formidable player in the AI domain, promising broader and more practical AI applications.

---

4. Hugging Face’s HuggingSnap iOS App

Introduction to HuggingSnap

The final segment of the episode highlights Hugging Face’s new application, HuggingSnap, an AI-powered vision app designed for iOS devices. This app distinguishes itself by performing all processing offline, enhancing privacy and usability.

A [12:09]: "We looked at Hugging Face's new app Hugging Snap, which is a great example of how AI can be used in a practical and privacy-conscious way."

Features and Offline Processing

HuggingSnap leverages Hugging Face’s vision model, Smolv LM2, to provide real-time image recognition and description without relying on cloud-based processing.

B [12:26]: "It's like having a personal AI assistant that can see what you're seeing."

Users can point their iPhones at objects, scenes, or text, and receive immediate, accurate information, all while maintaining data privacy since no information is transmitted over the internet.

A [12:35]: "You can point your phone at something and ask it questions like what is this? Or what does this sign say?"

Privacy and Practical Advantages

By operating offline, HuggingSnap offers enhanced privacy, energy efficiency, and usability in areas with limited or no internet connectivity.

A [12:47]: "Hugging Snap is different because it's all offline and that means it's more private, it's more energy efficient, and it doesn't rely on an Internet connection."

Applications and Compatibility

The app is versatile, suitable for shopping, traveling, learning, and exploring environments. It is compatible with iPhones running iOS 18 or later, as well as macOS and Apple Vision Pro devices.

B [13:07]: "Absolutely. And it's available for iPhones running iOS 18 or later. And it's also compatible with macOS and Apple Vision Pro."

Conclusion

HuggingSnap exemplifies the practical application of AI in everyday life, balancing functionality with user privacy. Its offline processing capability sets it apart from competitors, making it a valuable tool for a wide range of users seeking secure and efficient AI-powered vision assistance.

---

Final Thoughts and Future Outlook

In wrapping up the episode, Hosts A and B reflect on the rapid advancements and diverse directions in AI development. They emphasize the dual focus on increasing AI capabilities while ensuring ethical responsibility and user privacy.

B [14:15]: "Yeah, it's amazing to see how fast this field is moving, and it really highlights some of the big themes we're seeing in AI right now. The constant drive for more power, the growing need for responsible development, the incredible advancements in reasoning capabilities, and this shift towards user privacy and control."

They acknowledge that AI is at a pivotal point, with future developments poised to significantly impact various aspects of daily life, scientific research, and technological innovation.

A [14:54]: "I can't wait to see what happens next."

Conclusion

This episode of AI Deep Dive offers an insightful examination of the latest AI models and applications, highlighting both groundbreaking advancements and critical challenges. From OpenAI’s powerful yet pricey O1 Pro and LG’s efficient EXAONE Deep to Hugging Face’s privacy-focused HuggingSnap app, the discussion provides a comprehensive overview of the current AI landscape. Additionally, the conversation on the privacy complaint against ChatGPT underscores the importance of ethical considerations in AI development. As AI continues to evolve rapidly, staying informed and critically assessing these developments will be essential for developers, enthusiasts, and the general public alike.

Transcript

[A]

Man, the AI world is like a rocket these days, huh? It's crazy how fast things are moving. You turn your head for a second, and boom, there's like, five new things to try and understand.

[B]

Yeah, I feel that keeping up with it all is. Well, it's a lot. We get bombarded with news every single day. New models, new breakthroughs. And sometimes it's hard to know what's really important, you know?

[A]

Totally. It's like trying to drink from a fire hose sometimes. So that's what we're doing here today, right? Taking a deep breath, stepping back, and focusing on what really matters in this flood of AI news.

[B]

Exactly. It's about getting a clear picture of what's happening and what it all means. So you've brought some really fascinating stuff to look at today.

[A]

Yeah, we've got some pretty big developments to unpack.

[B]

We've got OpenAI's new model. This one is called 01 Pro, and it's supposed to be, well, pretty amazing.

[A]

And there's this privacy complaint. Right. This one looks pretty serious. It's about AI making stuff up. And it's not just like, you know, getting someone's birthday wrong. We're talking about some potentially really damaging false information.

[B]

It's a big deal, for sure. And then we've got some really cool stuff, too. LG has stepped up their AI game in a big way with Exo one Deep. It's doing things that were, well, honestly, a little hard to believe until recently.

[A]

That's awesome. And we've got hugging face, always pushing the boundaries. They've released this new app, Hugging Snap, and get this, it uses AI to tell you what your phone camera is looking at, but it's all offline. All the processing happens right there on your phone.

[B]

Yeah, it's amazing.

[A]

So it sounds like we've got a lot to cover. We're looking at how powerful AI is becoming, but also how we need to use it responsibly and make sure it's accurate. And, of course, we've got some incredible advancements that could really change things.

[B]

Exactly. And, you know, that's what this deep dive is all about. Understanding these big trends and figuring out what they mean for all of us.

[A]

Okay, let's jump in. Let's start with OpenAI and this new O1 Pro model. What is this thing?

[B]

Well, they're calling it a more powerful version of their O1 model. Basically, it's like they took their already powerful O1 and gave it, you know, a turbo boost or something.

[A]

And this is aimed specifically at developers. Right. Developers who need a lot of computational power. Like a lot.

[B]

Exactly. And of course, that kind of power comes with a price. We're talking 150 per million dollar input tokens and $600 per million dollar output tokens. Wow.

[A]

That's way more expensive than the standard 01 or even their previous advanced model GPT 4.5.

[B]

It's a big jump for sure. And it kind of tells you they're aiming for a specific market here, developers who need the very best and are willing to pay for it. OpenAI seems to believe that the improvements are worth the extra cost.

[A]

So they're saying O1 Pro can handle those really tough problems better, and that's why it needs all that extra computing power.

[B]

That's what they're saying. They claim it came about because developers were asking for something that could handle the really hard stuff.

[A]

But here's where it gets interesting. You mentioned that some people who got early access to Zero1Pro through ChatGPT Pro, well, they weren't blown away. They were saying it struggled with things like Sudoku puzzles and even some optical illusion jokes.

[B]

Yeah, it's kind of weird, right? It makes you wonder, like, if it's so powerful, why is it having trouble with things that, you know, seem pretty basic to us.

[A]

Exactly. And then there's the whole question of the benchmarks. OpenAI's own tests showed it was only a little bit better than the regular O1 on things like coding and math, right?

[B]

A little bit better, yeah. Though they did say it was more reliable, so maybe that's the key. It's not necessarily a huge jump in overall power, but it's more consistent. And maybe for certain developers, that consistency is worth the extra money.

[A]

Okay, so for all of us listening out there, what's the takeaway here? What does this mean for us?

[B]

It's a good question. I think we're seeing this trend towards increasingly specialized AI tools. You know, in the future, it's probably going to be less about do I have AI? And more about which AI is the right tool for this specific job.

[A]

Makes sense. And it means we have to be smart about how we evaluate these tools. We need to look at real world performance, not just benchmarks or what the companies tell us. Does it actually deliver the improvements it promises? Is it worth the cost?

[B]

Absolutely. Those are the big questions.

[A]

Okay, so let's shift gears a little bit. We've been talking about how powerful AI is becoming, but there's also this whole other side to it, right? This question of responsibility of Using AI ethically and making sure this doesn't do any harm.

[B]

Right. And that brings us to this privacy complaint against ChatGPT. This one is pretty serious.

[A]

Yeah, this one caught my eye. It's about AI making things up. And it's not just harmless little errors. This time it accused someone of a very serious crime, something that could seriously damage someone's reputation.

[B]

Yeah. So there's this organization, noib. They're a privacy rights group, and they're representing a guy from Norway. Chatgpt accused him of, well, murdering his children, which, of course, he absolutely did not do.

[A]

Wow, that's intense. And it seems like a whole different level of error than, you know, getting someone's birth date wrong.

[B]

Right. This is about defamation. It's about AI spreading false information that could have a really devastating impact on someone's life.

[A]

And NOIB is arguing that OpenAI isn't doing enough to protect people from this kind of harm. They're saying that under gdpr, people have the right to correct inaccurate information about themselves. And OpenAI isn't making it easy to do that.

[B]

Right. Nove is saying a simple disclaimer like, hey, ChatGPT can make mistakes. That's not enough. They are saying OpenAI is legally obligated to make sure the information they're putting out there is accurate. And if it's not, people need a way to fix it.

[A]

And the potential penalties under GDPR are pretty huge, right? Like 4% of a company's global revenue.

[B]

Yeah. It's a big deal. And it's not just theoretical. Remember what happened in Italy last year? They temporarily blocked ChatGPT because of GDPR concerns, and then they hit OpenAI with a fine. So European regulators are definitely not messing around.

[A]

But it's also interesting that other regulators, like in Ireland and Poland, they seem to be taking a more cautious approach. They haven't decided how GDPR applies to generative AI yet. So it sounds like this complaint could drag on for a while.

[B]

It could, yeah. It seems like NOIB is trying to push regulators to take this seriously. You know, they want them to understand the danger of these AI hallucinations, where the AI just makes stuff up and.

[A]

They'Re right to be concerned. You mentioned some pretty specific details about this case. ChatGPT got some basic facts right, like the number of children, the guy had, their genders in his hometown, but then it just completely fabricated this horrific story about him murdering them. That's. That's really creepy.

[B]

Yeah, it's unsettling. It kind of makes it worse. Right. Because Some of the information is right. So it almost makes the lie more believable.

[A]

Exactly. And there's this theory that maybe the AI is picking up on patterns in the data it was trained on. Like maybe there were stor about parents harming their children in the training data.

[B]

That's possible. These language models learn from massive amounts of text, so it's not hard to imagine that some of that text could be pretty dark. And if the AI sees those patterns often enough, maybe it starts to think it's normal or something.

[A]

It's a chilling thought, but at least ChatGPT stops saying those specific things about the guy after an update. Right? You said maybe it's because now it checks the Internet before talking about real people.

[B]

Yeah, that's the theory.

[A]

Still, it's kind of disturbing to think that even if ChatGPT isn't saying those things out loud anymore, that information might still be in there somewhere in its memory.

[B]

That's a good point. NOIB is definitely concerned about that. They're saying that even if OpenAI tries to hide the false information from users, it still needs to be corrected. Like it can't just pretend it never happened.

[A]

Right, so this complaint was filed in Norway. Right. But OpenAI's European headquarters are in Ireland, so it'll be interesting to see how that plays out.

[B]

Yeah, well, NOIB is hoping the Norwegian authorities will take this seriously, but it's not clear if they will.

[A]

And it reminds me of that case in Austria, the one NOIB back there. It was referred to Ireland, and it's still unresolved. So it seems like these GDPR complaints against big tech companies, they can take a long time to play out.

[B]

Yeah, the wheels of justice turn slowly, especially when you're dealing with global companies and complex regulations.

[A]

So what does all this mean for everyone listening?

[B]

Well, I think it highlights how important data privacy is in the age of AI. This is about protecting ourselves, but it's also about protecting others from being harmed by inaccurate information. We need AI that's accurate, transparent, and accountable.

[A]

I agree. Okay, let's move on to something a little less heavy. It sounds like LG has really stepped up their AI game with this X01 deep model.

[B]

Yeah, they're making some pretty bold claims. They're positioning Exon Deep as a new standard for reasoning AI, and they're taking on some of the big players in the field.

[A]

You mentioned agentic AI earlier. What exactly does that mean?

[B]

Well, it's the idea that AI systems can be more than just pattern recognizers. It's about AI that can actually think for itself, that can make decisions and solve problems without needing a human to hold its hand every step of the way. And LG thinks EXO and DEEP is a big step in that direction.

[A]

So they're saying it's really good at logic, at scientific reasoning and at coding, basically all the stuff that requires. Well, a pretty smart brain.

[B]

That's right. And they've got some impressive results to back up those claims.

[A]

Yeah. You mentioned their performance on math benchmarks was particularly impressive. All three of their models, the 32 billion parameter 1, the 7.8 billion 1, and even the smaller 2.4 billion 1, they all got the highest possible score on the CSAT 2025 mathematics section.

[B]

It's pretty amazing. That's a major college entran exam in South Korea, so it's not exactly walk in the Park.

[A]

And the 32 billion parameter model, it also scored really high on the AME. That's the American Invitational Mathematics Examination. It got a 90.0 on the 2024 version and it was basically performing as well as DeepSeek R1 on the 2025 version.

[B]

Which is crazy because DeepSeek R1 is way bigger. It has 671 billion parameters. So Exxon Deep is doing more with less.

[A]

Yeah, that's really efficient. It's like they've managed to pack a lot of brain power into a smaller package. And it's not just the big model. Right. The smaller 7.8 billion and 2.4 billion parameter models, they also did really well on those benchmarks. In fact, the 2.4 billion model is one of the best performing models for edge devices, according to Hugging Faces LLM Leaderboard.

[B]

It's pretty impressive. It seems like LG is focusing on making AI that's not just powerful, but also accessible. You know, AI that can run on devices that are more affordable and easier to use.

[A]

Right. So they're not just chasing bigger and bigger models, they're also trying to make AI more practical.

[B]

Exactly. And their science encoding performance is also worth noting. They say Exa Bundeep showed competitive superiority on tests like GPQA diamond, which is for PhD level science problems, and Live Code Bench, which is for coding.

[A]

So they're saying it's better than other models of similar size in those areas.

[B]

That's what they're saying.

[A]

And their general language understanding is also strong. The 32 billion parameter model got the highest MMLU score of any Korean model.

[B]

It's a really well rounded model.

[A]

Yeah.

[B]

It's not just good at one thing, it seems to Be good at, well, a lot of things.

[A]

And it's interesting that Epoch AI, a pretty well respected organization in the AI field, has listed both EXO1 deep 32B and the previous EXOEN 3.5 as notable AI models. And they're the only Korean models to get that recognition in the past two years.

[B]

That's a pretty big endorsement. It suggests that LG is really doing something special here.

[A]

So what's the takeaway for our listeners? What does all this mean for them?

[B]

Well, I think it shows that the field of reasoning AI is advancing rapidly. We're seeing models that are more capable and efficient, which means we could see some really interesting AI applications in the future. Maybe AI that can help us with scientific discoveries or AI that can write better code.

[A]

That's exciting stuff. Okay, let's wrap up with Hugging Face's new app, Hugging Snap.

[B]

This one is really cool. It's basically an AI powered vision app for your iPhone.

[A]

And the big deal is that it all happens offline. Right. All the processing is done right there on your phone. So you don't have to worry about your data being sent to the cloud.

[B]

Exactly. It uses Hugging Face's own vision model, Smolv LM2. Yeah. And it's all self contained.

[A]

So it's like having a personal AI assistant that can see what you're seeing. You can point your phone at something and ask it questions like what is this? Or what does this sign say?

[B]

Yeah. It can identify objects, describe scenes, even read text.

[A]

Now it's true that there are other apps that do similar things and Apple is even working on its own AI powered features for iPhones.

[B]

Right, but Hugging Snap is different because it's all offline and that means it's.

[A]

More private, it's more energy efficient, and it doesn't rely on an Internet connection.

[B]

Exactly. You could use it anywhere, even if you're off the grid.

[A]

I can see this being really useful for all sorts of things. Shopping, traveling, learning, even just exploring your surroundings.

[B]

Absolutely. And it's available for iPhones running iOS 18 or later. And it's also compatible with macOS and Apple Vision Pro.

[A]

So it's using the latest and greatest technology from Apple.

[B]

That's right.

[A]

So for everyone listening, Hugging Snap is a great example of how AI can be used to make our lives easier and more convenient while still respecting our privacy.

[B]

And it shows that you don't always need a massive cloud infrastructure to do cool things with AI.

[A]

This has been a really interesting look at the latest developments in AI.

[B]

Yeah, we've covered a lot of ground.

[A]

We talked about OpenAI's new O1 Pro model, which is super powerful but also super expensive and maybe not quite as impressive as some people were hoping. We discussed that concerning privacy complaint against ChatGPT, which raises some really important questions about how we hold AI accountable for its mistakes. And then we got into some really exciting stuff like LG's X01 deep model, which seems to be pushing the boundaries of what AI can do in terms of reasoning and problem solving. And finally, we looked at Hugging Face's new app Hugging Snap, which is a great example of how AI can be used in a practical and privacy conscious way.

[B]

Yeah, it's amazing to see how fast this field is moving, and it really highlights some of the big themes we're seeing in AI right now. The constant drive for more power, the growing need for responsible development, the incredible advancements in reasoning capabilities, and this shift towards user privacy and control.

[A]

So, given everything we've talked about today, what do you think the future holds for AI? What are the developments you're most excited about? And how do you think AI will impact our lives in the years to come? It really feels like we're at a turning point.

[B]

Yeah, it does. It's hard to say exactly what the future holds, but one thing's for sure, it's going to be an interesting ride.

[A]

I can't wait to see what happens next.

[B]

Me neither.